ASI-LIB-019 16 Jun 2025 technical white paper

AlphaEvolve: A Coding Agent for Scientific and Algorithmic Discovery

Alexander Novikov, Ngan Vu, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J. R. Ruiz, Abbas Mehrabian, M. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Pushmeet Kohli, Matej Balog

Figure via ar5iv rendering of arXiv:2506.13131

AlphaEvolve is a high-signal demonstration of evaluator-grounded discovery. It uses LLMs to modify code and evolutionary selection to retain improvements against objective evaluators.

Results to watch

The paper reports improvements to data-center scheduling, hardware accelerator circuit design, training efficiency, and algorithmic discoveries including a procedure for 4 x 4 complex matrix multiplication using 48 scalar multiplications.

ASI relevance

This is the shape of practical machine intelligence compounding: systems that can improve the computational substrate used to train later systems, then feed those gains back into the next search.